OpenAI has introduced its new product, ChatGPT Agent, which expands the capabilities of ChatGPT. The company explained that this new tool goes beyond traditional chatbots and can complete complex, multi-step tasks autonomously, using a “virtual computer.”
ChatGPT Agent Transforms the User Experience
ChatGPT Agent Product Lead Yash Kumar and Research Lead Isa Fulford explained that the tool is powered by a custom-developed model. This model can analyze and summarize calendar meetings on behalf of the user, create a shopping list for a family breakfast, and even prepare presentations by analyzing data from competitor companies.
The model used by the tool has no specific name, but it combines the capabilities of two existing OpenAI tools, Operator and Deep Research. This system runs on a platform with multiple tools, including a text browser, image browser, and command-line browser, and users can also import their own data. All of these functions rely on an infrastructure trained using reinforcement learning to perform multi-stage tasks.
The Operator and Deep Research teams were combined to develop ChatGPT Agent. This combined team currently consists of 20 to 35 people from product and research areas.
A demo presentation during the launch showcased examples of the Agent connecting to Google Calendar to check available time slots and comparing restaurant reservations on OpenTable to make plans on the user’s behalf.
The user can intervene with additional information or change criteria during the process. It also demonstrated the ability to create detailed research reports on specific topics, such as “the rise of Labubus toys and Beanie Babies.”
Isa Fulford stated that using the Agent for tasks like shopping is much more effective and comprehensive than using the Operator alone. Yash Kumar stated that he uses the ChatGPT Agent to automate small tasks in his own life. He explained that he no longer manually performs routine tasks, such as creating parking requests at the OpenAI office in San Francisco every Thursday.
Kumar explained that the Agent runs not just on a browser but on a system that simulates an entire computer, allowing it to use much more advanced tools. However, the demo also demonstrated that the system can sometimes be slow in terms of speed.
Kumar stated that they focus on completing complex tasks accurately rather than speed. Fulford emphasized that such tasks can take hours to complete manually, and that the Agent’s 15-30 minute processing time represents a significant advantage.
The Agent doesn’t perform any “irreversible” actions without user approval during the process. For example, it asks for permission before sending an email or making a reservation. This ensures the system remains under user control.
OpenAI announced that the new model used for the Agent has higher capabilities, enabling special protections against biological and chemical abuse risks. The company states that there is no evidence that the model can directly produce dangerous results.
However, additional security layers have been implemented for high-risk scenarios. Similar measures were implemented during the launch of Anthropic’s Claude Opus 4 model.
The ChatGPT Agent is positioned as a step forward in demonstrating that AI is becoming more than just a tool to provide information, but rather an active digital assistant. What are your thoughts on this? Share your thoughts with us in the comments section below.
